This repo hosts multiple knowledge tracing/IRT prototypes. The active work is in mirt-dkvmn/, which implements a multi-dimensional DKVMN with a GPCM head and synthetic MIRT data generation.

Use the vrec-env conda environment for all runs:

source ~/anaconda3/etc/profile.d/conda.sh
conda activate vrec-env

Most scripts expect PYTHONPATH=mirt-dkvmn/src.

Generate a synthetic dataset (example: 5000 students, 1000 items, 5 categories, 3 traits):

PYTHONPATH=mirt-dkvmn/src python mirt-dkvmn/scripts/data_gen.py \
  --name synthetic_5000_1000_5_d3 --n_traits 3 --min_seq 120 --max_seq 150 \
  --output_dir mirt-dkvmn/data

Train with a config:

PYTHONPATH=mirt-dkvmn/src python mirt-dkvmn/scripts/train.py \
  --config mirt-dkvmn/configs/large_d3_opt3.yaml

Plot recovery/metrics:

PYTHONPATH=mirt-dkvmn/src python mirt-dkvmn/scripts/plot_recovery.py \
  --config mirt-dkvmn/configs/large_d3_opt3.yaml \
  --checkpoint mirt-dkvmn/artifacts/large_d3_opt3/last.pt \
  --output mirt-dkvmn/artifacts/large_d3_opt3/recovery_plots

python mirt-dkvmn/scripts/plot_metrics.py \
  --metrics mirt-dkvmn/artifacts/large_d3_opt3/metrics.csv \
  --output mirt-dkvmn/artifacts/large_d3_opt3/metric_plots

• mirt-dkvmn/: Current MIRT-DKVMN implementation, configs, data tools, plots.
• deep-gpcm/: Reference implementation and legacy scripts.
• dkvmn-ori/, dkvmn-torch/, akt/, deep-1pl/: Legacy or comparative baselines.
• updated_plan.tex: Math/architecture notes for the MIRT-DKVMN design.

• Synthetic datasets live under mirt-dkvmn/data/ and follow synthetic_<students>_<items>_<cats>_d<traits>.
• Training artifacts (metrics, checkpoints, plots) go in mirt-dkvmn/artifacts/.
• Older datasets and results are archived in mirt-dkvmn/archive/.

• GPU usage is controlled by base.device in each config; training falls back to CPU if CUDA is unavailable.
• Dataset regeneration should match the current GPCM formulation to keep recovery plots meaningful.